A statistical ensemble of large-eddy simulations (LES) is run simultaneously for the same flow. The information provided by the different large-scale velocity fields is used in an ensemble-averaged version of the dynamic model. This produces local model parameters that only depend on the statistical properties of the flow. An important property of the ensemble-averaged dynamic procedure is that it does not require any spatial averaging and can thus be used in fully inhomogeneous flows. Also, the ensemble of LES provides statistics of the large-scale velocity that can be used for building new models for the subgrid-scale stress tensor. The ensemble-averaged dynamic procedure has been implemented with various models for three flows: decaying isotropic turbulence, forced isotropic turbulence, and the time-developing plane wake. It is found that the results are almost independent of the number of LES in the statistical ensemble provided that the ensemble contains at least 16 realizations.